Menu Back to Poster-Presentations-Details

P15: Bayesian Prognostic Covariate Adjustment: a Method to Reduce Sample Sizes While Maintaining Power and Statistical Accuracy

Poster Presenter

David Walsh

Statistician
Unlearn.Ai
United States

Objectives

To demonstrate a novel Bayesian approach for analyzing clinical trials, which can produce accurate estimates of treatment effects with smaller sample sizes than traditional statistical methods.

Method

We trained a “prognostic model” on historic data from Alzheimer’s Disease (AD) trials to predict placebo outcomes for each subject in a study. These predictions were used in a Bayesian analysis both as covariates for each subject, and to construct a prior distribution for the mean placebo outcome.

Results

We proved that the proposed analysis controls the Type-I-error rate, when the variance of the prior distribution is tuned to the magnitude of any bias that the prognostic model may exhibit. In addition, we established how our Bayesian framework can be used to prospectively power a trial, i.e., to select the sample sizes such that the proposed analysis would be powered to detect a given effect size. We proved that in general, our method can achieve the same power as traditional analyses while requiring substantially smaller sample sizes. To assess the practical gains from our method, we re-analyzed a completed Phase II, double-blind, placebo-controlled trial in AD, focusing on three of the study’s secondary endpoints: Alzheimer's Disease Assessment Scale - Cognitive (ADAS-Cog11), Clinical Dementia Ratio - Sum of Boxes (CDR-SB), and Mini-Mental State Examination (MMSE). We tuned the prior variance through an assessment of the model’s performance on other clinical trials in AD with similar inclusion criteria. We performed a prospective power analysis for ADAS-Cog11; sampled the resulting sample sizes from the patient population in the original study; and applied our Bayesian method to the sampled data to obtain credible intervals for treatment effects on each of the endpoints. We found that the total sample size across the active treatment and placebo arms was reduced by 42%. The estimates of the treatment effects agreed closely with those of the original study for all three endpoints, with credible intervals which were no wider than the original confidence intervals.

Conclusion

Our Bayesian approach presents an opportunity to design and analyze clinical trials in a way that maintains statistical accuracy, while reducing the required sample size and the time and cost associated with recruiting and managing extra subjects. We have proven that our method has desirable operating characteristics in theory, and we have demonstrated the gains in practice. Our framework can be applied to any study where a prognostic model for placebo outcomes is available or possible, and it can produce accurate estimates of treatment effects across multiple endpoints simultaneously. This flexibility means that our methodology has the potential to radically improve the efficiency of clinical trials across many disease areas.